A report automatic generation method and system

Abstract

The application discloses a report automatic generation method and system, and relates to the technical field of artificial intelligence and natural language processing, which comprises the following steps: task analysis is performed on an initial query of a user to obtain a research intention, and the research intention is converted into a text framework draft based on a text hierarchical structure; the research intention is disassembled into a subtask set, and quality evaluation constraints of the subtasks are generated; information iterative retrieval is performed on the subtasks to obtain a global evidence pool, the global evidence pool is evaluated and verified based on the quality evaluation constraints, and a verified evidence set is obtained; the semantic-authoritative correlation degree between argumentation text in the text framework draft and the verified evidence set is used to judge the support degree type of the argumentation text, the argumentation text is purified and modified based on the support degree type, and a final research report is obtained. The application is used to solve the problems that the prior art lacks structured research quality evaluation, pre-task preparation and evidence purification in a report generation scene.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, and more specifically, to a method and system for automatic report generation. Background Technology

[0002] Currently, reports on technology selection and research typically serve high-risk business decisions directly, and their quality heavily relies on systematic retrieval, cross-validation, and structured synthesis of large-scale, heterogeneous, and significantly noisy network information. In practice, analysts often need to first clarify the research objectives, then continuously search for information and record key evidence around several sub-questions, and finally form the final report after multiple rounds of comparison and revision.

[0003] The existing technology has the following shortcomings:

[0004] (1) There is a lack of a pre-preparation stage for report writing, and a lack of pre-processing ability to transform vague user needs into structured research tasks and preliminary research results into a text framework that can be written.

[0005] (2) The lack of intermediate constraints and structured evaluation of research quality during the generation of research reports can easily lead to problems such as drift of research objectives and loose chain of evidence. Furthermore, the lack of an effective purification mechanism in the final draft generation stage can easily result in problems such as information piling up, loss of focus, and even inconsistency of facts.

[0006] To address the aforementioned issues, there is an urgent need for a method and system for automatic report generation. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an automatic report generation method and system to address the problems of the prior art lacking structured research quality assessment, preliminary task preparation, and evidence cleansing in report generation scenarios.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] The first aspect of this invention provides a method for automatically generating reports, comprising: parsing an initial user query to obtain a research intent, and converting the research intent into a text framework draft based on the text hierarchy structure; decomposing the research intent into a set of sub-tasks based on the content defects of the text framework draft, and generating quality assessment constraints for the sub-tasks; performing iterative information retrieval on the sub-tasks to obtain a global evidence pool, and evaluating and verifying the global evidence pool based on the quality assessment constraints to obtain a set of verified evidence; determining the support type of the argumentative text based on the semantic-authority correlation between the argumentative text in the text framework draft and the set of verified evidence, and purifying and correcting the argumentative text based on the support type to obtain the final research report.

[0010] In a preferred embodiment, the research intent is obtained by parsing the user's initial query, including: identifying the query task structure in the user's initial query based on a preset research task prompt, and generating the research intent based on the query task structure.

[0011] In a preferred embodiment, the research intent is transformed into a text framework draft based on the text hierarchy structure, including: extracting the text hierarchy structure of the research intent based on a preset draft generation prompt, and constructing a text framework draft based on the text hierarchy structure.

[0012] In a preferred embodiment, the research intent is decomposed into a set of subtasks based on the content defects of the text frame draft, including: obtaining the draft content defects of the text frame draft; decomposing the research intent into multiple subtasks according to the defect type of the draft content defects; calculating the priority score of the subtasks based on the alignment degree between the subtasks and the research intent; and setting the priority order of the subtasks according to the priority score; and constructing a set of subtasks according to the subtasks and the priority order.

[0013] In a preferred embodiment, the generation of quality assessment constraints for subtasks includes: introducing a quality constraint template that includes coverage constraints, accuracy constraints, and completeness constraints; obtaining task requirement parameters for the subtasks; determining the quality constraint template for the subtasks based on the parameter types of the task requirement parameters; and substituting the task requirement parameters into the quality constraint template to obtain the quality assessment constraints for the subtasks.

[0014] In a preferred embodiment, the global evidence pool is obtained by iteratively retrieving information from subtasks, including: generating a retrieval query target based on the subtasks; performing a global retrieval on the retrieval query target to obtain a retrieval result set; and filtering, deduplicating, and sorting the retrieval result set according to relevance to obtain the global evidence pool.

[0015] In a preferred embodiment, the global evidence pool is evaluated and verified based on quality assessment constraints to obtain a set of verified evidence, including: obtaining quality assessment dimensions of the global evidence pool based on quality assessment constraints; wherein, the quality assessment dimensions include content coverage, accuracy of key facts, and completeness of argumentation; evaluating and verifying the global evidence pool according to the quality assessment dimensions to obtain verified evidence, and constructing a set of verified evidence based on the verified evidence.

[0016] In a preferred embodiment, determining the support type of the argument text based on the semantic-authoritative correlation between the argument text of the text framework draft and the set of verified evidence includes: traversing the argument text of the text framework draft and calculating the semantic similarity between the argument text and the verified evidence in the set of verified evidence; obtaining the evidence channel type of the verified evidence and obtaining the authority weight value of the verified evidence based on the evidence channel type; calculating the semantic-authoritative correlation based on the semantic similarity and authority weight value, and determining the support type of the argument text based on the correlation, wherein the support type includes sufficient support, partial support, no support, and conflicting support.

[0017] In a preferred embodiment, the argument text is cleaned and revised based on the support level type to obtain the final research report, including: generating a cleansing and revision instruction based on the support level type, and performing evidence revision and evidence filling on the argument text according to the cleansing and revision instruction to obtain a cleaned text draft; and generating the final research report based on the cleaned text draft.

[0018] A second aspect of this invention provides an automatic report generation system, comprising: a draft generation module for parsing an initial user query to obtain research intent, and converting the research intent into a text framework draft based on a text hierarchical structure; a task planning module for decomposing the research intent into a set of sub-tasks based on the content defects of the text framework draft, and generating quality assessment constraints for the sub-tasks; a retrieval and execution module for iteratively retrieving information from the sub-tasks to obtain a global evidence pool, evaluating and verifying the global evidence pool based on the quality assessment constraints, and obtaining a set of verified evidence; and a purification and correction module for determining the support type of the argumentative text based on the semantic-authority correlation between the argumentative text in the text framework draft and the set of verified evidence, and purifying and correcting the argumentative text based on the support type to obtain the final research report.

[0019] The beneficial effects of this invention are:

[0020] 1. This invention obtains the research intent by parsing the user's initial query and transforms the research intent into a text framework draft based on the text hierarchy structure. Through these two preliminary stages of research intent and text framework draft generation, clear task boundaries and text framework support are provided for subsequent research, solving the problem of vague user needs and difficulty in directly initiating retrieval in research tasks.

[0021] 2. This invention, through a collaborative generation mechanism of subtask planning and iterative information retrieval, as well as quality assessment constraints, enables each subtask to have accurate evidence criteria, achieves fine-grained control over intermediate research results, and improves the quality of the final research report by purifying and revising the argumentative text of the text framework draft before the final research report is generated through the support level type. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a method for automatically generating reports.

[0023] Figure 2 A schematic diagram of the method for generating a set of subtasks based on a textual framework draft, provided in an embodiment of the present invention;

[0024] Figure 3 This is a schematic diagram of a method for evaluating and verifying a global evidence pool based on quality assessment constraints to obtain a set of verified evidence, provided in an embodiment of the present invention.

[0025] Figure 4 A schematic diagram of the method for determining the semantic-authoritative correlation between argumentative text based on a text framework draft and a set of verified evidence, provided for embodiments of the present invention;

[0026] Figure 5 This is a schematic diagram of the structure of an automatic report generation system. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0028] Example 1

[0029] Please see Figure 1 , Figure 1 This invention presents an automatic report generation method, comprising the following steps: parsing the user's initial query to obtain the research intent, and transforming the research intent into a text framework draft based on the text hierarchy structure; decomposing the research intent into a set of sub-tasks based on the content defects of the text framework draft, and generating quality assessment constraints for the sub-tasks; performing iterative information retrieval on the sub-tasks to obtain a global evidence pool, and evaluating and verifying the global evidence pool based on the quality assessment constraints to obtain a set of verified evidence; determining the support type of the argumentative text based on the semantic-authority correlation between the argumentative text in the text framework draft and the set of verified evidence, and purifying and correcting the argumentative text based on the support type to obtain the final research report.

[0030] S1. The initial user query is parsed to obtain the research intent. Based on the text structure of the intent, a draft text framework is generated, including:

[0031] S11. Perform task parsing on the user's initial query to obtain research intent, including: performing task parsing on the user's initial query based on preset research task prompts, identifying the query task structure, and generating research intent based on the query task structure. Specifically, based on the preset research prompts, obtain the preset task structure in the user's initial query, such as the research topic, research scope, core focus, potential search direction, and task constraints. This can be achieved by providing the user's initial query... Through function Research Intent ,in This indicates a research task prompt, and the function can be implemented based on a research intent clarification interface, whose input parameter is the user's initial query. and research task prompts The output of this interface is for research purposes. This includes fields such as research topic, research scope, core focus, potential search directions, and task constraints. Additionally, the preset research task prompts include hints on the research topic, research scope, core focus, potential search directions, and task constraints.

[0032] S12. Based on the text hierarchy structure, the research intent is transformed into a text framework draft, including: extracting the text hierarchy structure of the research intent based on preset draft generation prompts, and constructing a text framework draft based on the text hierarchy structure. Specifically, the text framework draft generation process is defined as a function. ,in Hints for generating drafts This represents the generated text framework draft. This function can be implemented through a draft generation interface, whose input parameter is the research intent. Draft generation tips The output is a text frame draft. Text frame draft Specifically, it includes fields such as text structure rules, logical framework templates, and content white space rules. The text structure rules include headings, chapters, and paragraph levels; the logical framework templates include an introduction-analysis-conclusion logical framework; and the content white space rules include pre-defined spaces for evidence. The preset draft generation prompts include prompts for the text structure rules, logical framework templates, and content white space rules.

[0033] S2. Based on the content defects of the text framework draft, the research intent is broken down into a set of sub-tasks, and quality assessment constraints for the sub-tasks are generated, including:

[0034] Please see Figure 2 S21. Based on the argumentative text of the text framework draft, generate a set of subtasks through task planning, as follows:

[0035] S211. Based on the research intent, the draft content defects of the text framework draft are verified. Specifically, the types of draft content defects include: (1) structural defects, which are verified chapter by chapter in the text framework draft. (1) The hierarchical integrity is verified by identifying missing chapters, logical gaps (such as only having an introduction framework without analysis and conclusion chapters) or parts with chaotic hierarchical structure, which is used to clarify the core direction of the structure improvement task; (2) The content defect type is verified by identifying the unfilled core arguments, missing data support, and unmarked citation positions for each paragraph outline, and classifying and statistically analyzing the defect type (data type, opinion type, citation type) and corresponding position, which is used to clarify the specific requirements of the content filling task; (3) The quality defect type is verified by combining the research intention. The top-level requirements, compared with the text framework draft The existing framework identifies parts that do not conform to the research objectives (such as the research intent requiring "coverage of data from the last 5 years," while the text framework draft...). The verification (without reserved placeholders for related data) is used to clarify the correction direction for quality optimization tasks.

[0036] S212. Based on the type of defects in the draft content, the research intention is broken down into multiple sub-tasks. Specifically, the sub-tasks are divided into three categories according to the type of defects: (1) Structural improvement sub-tasks: focusing on the structural defects of the text framework draft, such as "supplementing the missing 'Conclusions and Suggestions' chapter framework in the text framework draft" and "adjusting the hierarchical logic of a certain chapter to ensure coherence"; (2) Content filling sub-tasks: focusing on the content defects of the text framework draft, such as "supplementing the 'Market Size' paragraph with authoritative data and statistical standards for 2025" and "supplementing a certain core argument with more than two authoritative supporting documents"; (3) Quality optimization sub-tasks: focusing on the quality defects of the text framework draft, such as "verifying the accuracy of a certain data in the text framework draft and correcting the erroneous values" and "unifying the citation format of the text framework draft to ensure standardization".

[0037] S213. Based on the alignment degree between subtasks and research intentions, calculate the priority score of each subtask, and set the priority ranking of subtasks according to the priority score. Specifically, the alignment degree is reflected by semantic matching degree and target contribution. A task semantic vector is generated based on the task text description of the subtask, and an intention semantic vector is generated based on the structured text of the research intention. The semantic matching degree between the subtask and the research intention is calculated based on the task semantic vector and the intention semantic vector. The calculation function is as follows:

[0038]

[0039] In the formula, T represents the subtask obtained from the decomposition. For task semantic vectors; As the intent semantic vector, based on semantic matching degree It reflects the semantic matching degree between the subtask and the research intention, and the value range is [-1,1]. After normalization, it will be mapped to the interval [0,1].

[0040] Structured text extraction of research objectives based on research intent , where g i Let i represent the i-th specific research objective in the research objective set G, and The contribution of subtasks to the research objectives of the research objective set is evaluated. If a subtask makes no contribution to the research objective, its contribution to the research objective is 0; if a subtask makes a core contribution to the research objective, its contribution to the research objective is 1; and if a subtask makes a secondary contribution to the research objective, its contribution to the research objective is 0.5. The contribution of subtasks to the research objective B is calculated using the following function. :

[0041]

[0042] In the formula, The contribution of a subtask to each sub-goal. The number of research objectives in the research objective set;

[0043] Calculate the priority score of the subtask using the following function. :

[0044]

[0045] In the formula, This represents the semantic similarity weight, which can be 0.6. The subtasks are prioritized based on their priority scores, from highest to lowest.

[0046] S214. Based on the subtasks and priorities, construct a set of subtasks. Specifically, sort and combine the subtasks based on their priorities to obtain the set of subtasks.

[0047] S22. Quality evaluation constraints for generated subtasks, including:

[0048] S221. Introduce a quality constraint template that includes coverage constraints, accuracy constraints, and completeness constraints. Obtain task requirement parameters based on the task text description of the subtask, and determine the quality constraint template of the subtask based on the parameter types of the task requirement parameters. Specifically, if the task requirement parameters of the subtask include requirements for full coverage of scope, scenario, and data source, then the quality constraint template of the subtask includes coverage constraints; if the task requirement parameters of the subtask include requirements for data accuracy, factual authenticity, error control, and authoritative sources, then the quality constraint template of the subtask includes accuracy constraints; if the task requirement parameters of the subtask include requirements for required fields, key points, and detailed dimensions, then the quality constraint template of the subtask includes completeness constraints.

[0049] S222. Substitute the task requirement parameters into the quality constraint template to obtain the quality assessment constraints for the subtask. Specifically, substitute the task requirement parameters in the subtask into the various constraints of its corresponding quality constraint template. Substitute the task requirement parameters that include full coverage of scope, scenario, and data source into the coverage constraint; substitute the task requirement parameters that include data accuracy, factual authenticity, error control, and authoritative source requirements into the accuracy constraint; and substitute the task requirement parameters that include essential fields, key points, and detailed dimensions into the integrity constraint. Combine the coverage constraint, accuracy constraint, and integrity constraint to construct the quality assessment constraints for the subtask.

[0050] S3. Perform iterative information retrieval on the subtasks to obtain a global evidence pool. Evaluate and verify the global evidence pool based on quality assessment constraints to obtain a set of verified evidence, including:

[0051] S31. Based on the subtasks and their context, generate a retrieval query set and call tools (such as search engines / databases / knowledge bases) to perform a retrieval on the query set to obtain the original result set; after filtering, deduplicating, and sorting the original result set by relevance, supplement the original result set to the global evidence pool, including:

[0052] S311. Filter the original result set. Specifically, a multi-source filtering dimension can be preset, including relevance screening, source authority screening, timeliness filtering, and completeness filtering. Search results in the original result set that do not meet the multi-source filtering dimension are initially screened and removed. Among them, relevance screening can remove completely irrelevant search results based on the semantic matching degree between the result text and the search query; source authority screening can remove search results from low-quality / unreliable sources based on the authority level of the data source; timeliness filtering can remove invalid search results that are beyond the time range specified by the task or have been published too long; completeness filtering can remove search results with incomplete content, missing key information, or damaged format.

[0053] S312. Deduplicatize the original result set. Specifically, generate SimHash fingerprints based on the filtered search results and calculate the Hamming distance between the SimHash fingerprints of any two search results. If the Hamming distance is less than a preset distance threshold (e.g., ≤3), the two search results are determined to be highly similar, and only one search result is retained (prioritizing the version with higher authority).

[0054] S313. The original result set is ranked by relevance. Specifically, this is done by normalizing the cosine similarity between the search result vector and the search query vector to obtain a semantic relevance score; an authority score is obtained based on the source channel of the search results: if the source channel is an authoritative journal or official data, the authority score is 1; if the source channel is ordinary media, the authority score is 0.7; if the source channel is an unreliable channel, the authority score is 0.3; furthermore, based on the current time, results closer to the current time have higher timeliness scores, calculated using the following function. The specific formula is as follows:

[0055]

[0056] In the formula, The time difference between the result release time and the current time (in days); This is a decay constant (which can be adjusted according to the task type; tasks with high timeliness requirements should take a larger value).

[0057] Based on preset score weights, the semantic relevance score, authority score, and timeliness score are weighted and calculated to obtain the relevance score. The score weights include voice weight, authority weight, and timeliness weight. The specific calculation formula is as follows:

[0058]

[0059] In the formula, The semantic relevance score, This is the semantic weight, with a default value of 0.5; For authority score, This is the authority weight, with a default value of 0.3; For timeliness score, This is the timeliness weight, with a preset value of 0.2. The search results are sorted by relevance score from highest to lowest to obtain the processed original result set.

[0060] Please see Figure 3S32. Evaluate and verify the global evidence pool based on quality assessment constraints to obtain a set of verified evidence, including:

[0061] S321. Obtain the quality assessment dimensions of evidence in the global evidence pool based on quality assessment constraints. Specifically, the quality assessment dimensions include content coverage, accuracy of key facts, and completeness of argumentation. The content coverage of the evidence is used to assess whether the required information for the sub-task is sufficiently covered. (0.0-1.0) Quantitative representation, assessing the accuracy of factual statements through the accuracy of key facts in the evidence, in order to (0.0-1.0) Quantitative representation, assessing the completeness of the argument chain through the completeness of evidence, in order to (0.0-1.0) The specific calculation process for content coverage, key fact accuracy, and argument completeness, expressed quantitatively, is as follows:

[0062] (1) Construct a standard coverage dimension set, which may include time dimension, region dimension, subdivided scenario dimension, etc. Extract the covered dimension set based on the information iterative retrieval results, and calculate the content coverage using the following calculation formula. :

[0063]

[0064] In the formula, The number of dimensions in the standard coverage dimension set. This represents the number of dimensions in the covered dimension set.

[0065] (2) Construct a list of key facts Each fact is compared to authoritative data. If a fact is completely consistent with the authoritative data source, its score is 1; if there is a deviation within a preset range, its score is 0.5-0.9; if there are obvious errors or contradictions, its score is 0. The accuracy of key facts is calculated using the following function. :

[0066]

[0067] In the formula, This refers to the number of facts in the key fact list; Numbering of facts; The score is the score for the fact.

[0068] (3) Predefine the set of essential argument elements, which may include argument elements such as preconditions, causal analysis, current data, influencing factors, conclusions and recommendations, and risk warnings. Extract the argument elements already available in the information iterative retrieval results and calculate the argument completeness using the following formula. :

[0069]

[0070] In the formula, The number of elements in the essential argument set. To achieve the required number of elements for argumentation.

[0071] S322. Determine the verification status of the evidence to be verified in the global evidence pool based on the quality assessment dimensions. Specifically, the verification status includes qualified, needing supplementation, needing correction, narrowing the scope, and irrecoverable. The verification status is determined based on a comprehensive set of pre-defined rules, considering factors such as content coverage, accuracy of key facts, and completeness of argumentation within the quality assessment dimensions. For example, the comprehensive rules may include Rule 1, Rule 2, and Rule 3, where Rule 1 is for when the accuracy of key facts... The status is determined to be pending supplementation; Rule 2 is when the content coverage... It is judged to be in a qualified state when the content coverage is satisfactory. It is determined to be in a state of needing to be supplemented, when the content coverage is... The scope is narrowed; Rule 3 is when the completeness of the argument is... This is determined to be in a state requiring supplementation. The preset content coverage threshold is defined as follows. Rule 1, Rule 2, and Rule 3 are parallel and independent judgment rules. Their logical triggering relationship is that as long as any one rule is met, the corresponding verification judgment state will be triggered. When multiple rules are met simultaneously and different verification judgment states are triggered, the state with the highest strictness will be used. The strictness is ordered from high to low as follows: irrecoverable state, narrowed scope state, state to be corrected, state to be supplemented, and qualified state.

[0072] S323. Evaluate and verify the global evidence pool based on the quality assessment dimensions to obtain verified evidence, and construct a set of verified evidence based on the verified evidence. Specifically: when the verification judgment status of the evidence to be verified is qualified, the evidence to be verified is recorded as verified data; when the verification judgment status of the evidence to be verified is to be supplemented, the evidence to be verified is recorded as verified data, and the search-verification loop is re-executed; when the verification judgment status of the evidence to be verified is to be corrected, a clearer search query target is regenerated based on the sub-task, and the original quality assessment constraints are inherited before re-verification; when the verification judgment status of the evidence to be verified is to narrow the scope, a more specific and searchable search query target is regenerated based on the sub-task, and re-verification is performed; when the verification judgment status of the evidence to be verified is irrecoverable, the current research content is terminated, the sub-task set and quality assessment constraints are updated, and evidence search and verification are performed again.

[0073] S4. Based on the semantic-authority correlation between the argumentative text in the draft text framework and the verified evidence set, the support level of the argumentative text is determined. The argumentative text is then refined and corrected based on this support level, resulting in the final research report, including:

[0074] Please see Figure 4 S41. The semantic-authoritative relevance of the argumentative text based on the draft text framework to the set of verified evidence is as follows:

[0075] S411. Traverse the argument text of the text framework draft and calculate the semantic similarity between the argument text and the verified evidence in the verified evidence set. Specifically, generate argument semantic vectors based on the argument texts, generate evidence semantic vectors based on the verified evidence, and calculate the semantic similarity using the following calculation function. :

[0076]

[0077] In the formula, Let c be the semantic vector of the text, with dimension n and components n. ; For the i-th piece of evidence The semantic vector, with dimension n and components. ; represents the dot product of two vectors; k is the dimension number.

[0078] S412. Obtain the evidence channel type for verification evidence, and obtain the authority weight value of the verification evidence based on the evidence channel type. Specifically, when the evidence channel type of the verification evidence is an authoritative journal or official data, the authority weight value of the evidence is 1; when the evidence channel type of the verification evidence is ordinary media, the authority weight value of the evidence is 0.7; when the evidence channel type of the verification evidence is an informal channel, the authority weight value of the evidence is 0.3.

[0079] S413. Semantic-authority relevance between the argumentative text of the draft text framework and the set of verified evidence. The calculation formula is as follows:

[0080]

[0081] In the formula, For argumentative texts, This is the i-th piece of evidence in the global evidence pool; For semantic similarity; This represents the authority weight of the evidence. This is the similarity weight, and can be set to 0.7.

[0082] S414. Based on relevance, the argument text is classified into support types including fully supported, partially supported, unsupported, and conflicting supported. Specifically, the top K pieces of evidence with the highest semantic similarity (K=5 by default) are selected as candidate evidence, and a candidate evidence set is constructed for the argument text. When there are ≥3 candidate pieces of evidence in the candidate evidence set and the semantic-authority relevance is ≥0.8, the support type of the argument text is determined to be fully supported. When there are ≥2 candidate pieces of evidence in the candidate evidence set and the semantic-authority relevance is in the interval [0.5, 0.8), the support type of the argument text is determined to be partially supported. When there are no candidate pieces of evidence in the candidate evidence set and the semantic-authority relevance is <0.5, the support type of the argument text is determined to be unsupported. When there are two pieces of highly similar evidence in the candidate evidence set, one supporting the argument and the other opposing the argument, the support type of the argument text is determined to be conflicting supported.

[0083] S42. Based on the support level type, the argumentative text is purified and revised to obtain the final research report, including:

[0084] S421. Generate purification and correction instructions based on the support level type. Specifically, when the support level type of the argument text is "sufficiently supported," the purification and correction instructions for the argument text are to merge the argument sections with duplicate and redundant content; when the support level type of the argument text is "partially supported," the purification and correction instructions for the argument text are to delete the argument sections in the argument text that are unsupported and cannot be supplemented; when the support level type of the argument text is "unsupported," the purification and correction instructions for the argument text are to supplement the argument text with argument text content that has indirect supporting evidence; when the support level type of the argument text is "conflicting support," the purification and correction instructions for the argument text are to correct the argument sections in the argument text that contradict the evidence.

[0085] S422. Based on the purification and correction instructions, the argument text is modified and filled with evidence to obtain a purified text draft. Specifically, the evidence is modified based on the purification and correction instructions, and the modified evidence is filled into the argument text of the text framework draft to generate a purified text draft. The final research report is generated based on the purified text draft.

[0086] Example 2

[0087] Please see Figure 5 This paper presents an automatic report generation system, comprising: a draft generation module, used to parse the user's initial query to obtain the research intent, and to transform the research intent into a text framework draft based on the text hierarchy structure; a task planning module, used to decompose the research intent into a set of sub-tasks based on the content defects of the text framework draft, and to generate quality assessment constraints for the sub-tasks; a retrieval and execution module, used to perform iterative information retrieval on the sub-tasks to obtain a global evidence pool, and to evaluate and verify the global evidence pool based on the quality assessment constraints to obtain a set of verified evidence; and a cleansing and correction module, used to determine the support type of the argument text based on the semantic-authority correlation between the argument text in the text framework draft and the set of verified evidence, and to clean and correct the argument text based on the support type to obtain the final research report.

[0088] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0089] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0090] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0091] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0092] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0093] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for automatically generating reports, characterized in that, Includes the following steps: The initial user query is parsed to obtain the research intent, and the research intent is transformed into a text framework draft based on the text hierarchy structure. Based on the content defects of the text framework draft, the research intent is decomposed into a set of sub-tasks, and quality assessment constraints for the sub-tasks are generated. Information iterative retrieval is performed on subtasks to obtain a global evidence pool. The global evidence pool is then evaluated and verified based on quality assessment constraints to obtain a set of verified evidence. Based on the semantic-authoritative correlation between the argumentative text and the verified evidence set in the text framework draft, the support type of the argumentative text is determined, and the argumentative text is cleaned and revised based on the support type to obtain the final research report. Based on the semantic-authoritative correlation between the argumentative text in the draft text framework and the verified evidence set, determine the type of support for the argumentative text, including: Iterate through the argument text of the text framework draft and calculate the semantic similarity between the argument text and the verified evidence in the verified evidence set; The type of evidence channel used to obtain verification evidence is determined, and the authority weight value of the verification evidence is obtained based on the evidence channel type. The semantic-authority correlation degree is calculated based on the semantic similarity and authority weight values. The first K pieces of verification evidence are selected based on the similarity to form a candidate evidence set. When the number of candidate evidence in the candidate evidence set is ≥3 and the semantic-authority correlation is ≥0.8, the support type is sufficient support. When the number of candidate evidence is ≥2 and the semantic-authority correlation is in the interval [0.5, 0.8), the support type is partial support. When the candidate evidence set is empty and the semantic-authority correlation is <0.5, the support type is no support. When there are two pieces of highly similar evidence with opposite polarities in the candidate evidence set, the support type is conflicting support. The semantic-authority correlation The calculation formula is as follows: ； In the formula, For argumentative texts, This is the i-th piece of evidence in the global evidence pool; For semantic similarity; This represents the authority weight of the evidence. Similarity weights; Let c be the semantic vector of the text. For the i-th piece of evidence The semantic vector.

2. The automatic report generation method according to claim 1, characterized in that, The research intent is obtained by parsing the user's initial query, including: Based on the preset research task prompts, the system identifies the query task structure in the user's initial query and generates research intent based on the query task structure.

3. The automatic report generation method according to claim 1, characterized in that, Based on the hierarchical text structure, the research intent is transformed into a text framework draft, including: Based on the preset draft generation prompts, the text hierarchical structure of the research intention is extracted, and a text framework draft is constructed according to the text hierarchical structure, combined with the preset logical framework template and content white space rules. The text hierarchy includes headings, chapters, and paragraphs, as determined by text structure rules.

4. The automatic report generation method according to claim 1, characterized in that, Based on the content deficiencies of the text-frame draft, the research intent is broken down into a set of sub-tasks, including: Obtain the defects in the text framework draft and break down the research intent into multiple sub-tasks based on the defect types in the draft content. Based on the alignment between subtasks and research objectives, priority scores are calculated for subtasks, and priority ranking of subtasks is set according to priority scores. Based on the subtasks and their priority, construct a set of subtasks.

5. The automatic report generation method according to claim 1, characterized in that, The quality assessment constraints for the generated subtasks include: Introduce a quality constraint template that includes coverage constraints, accuracy constraints, and completeness constraints; Obtain the task requirement parameters of the subtask, and determine the quality constraint template of the subtask based on the parameter types of the task requirement parameters; Substituting the task requirement parameters into the quality constraint template yields the quality assessment constraints for the subtasks.

6. The automatic report generation method according to claim 1, characterized in that, The global evidence pool is obtained by iteratively retrieving information from subtasks, including: Based on subtasks, a search query target is generated, and a global search is performed on the search query target to obtain a search result set; The search result set is filtered, deduplicated, and sorted by relevance to obtain a global evidence pool.

7. The automatic report generation method according to claim 1, characterized in that, The global evidence pool is evaluated and verified based on quality assessment constraints, resulting in a set of verified evidence, including: The quality assessment dimensions of the global evidence pool are obtained based on quality assessment constraints; these dimensions include content coverage, accuracy of key facts, and completeness of argumentation. The global evidence pool is evaluated and verified based on the aforementioned quality assessment dimensions to obtain verified evidence, and a set of verified evidence is constructed based on the verified evidence.

8. The automatic report generation method according to claim 1, characterized in that, The argumentative text was purified and revised based on the support level type, resulting in the final research report, including: Based on the support level type, a cleanup and correction instruction is generated, and the argument text is corrected and filled with evidence according to the cleanup and correction instruction to obtain a cleaned text draft. The final research report is generated based on the purified text draft.

9. A report automatic generation system, characterized in that, include: The draft generation module is used to parse the user's initial query to obtain the research intent, and to convert the research intent into a text framework draft based on the text hierarchy structure. The task planning module is used to break down research intentions into a set of sub-tasks based on the content defects of the text framework draft, and generate quality assessment constraints for the sub-tasks. The retrieval and execution module is used to perform iterative information retrieval on subtasks to obtain a global evidence pool, and to evaluate and verify the global evidence pool based on quality assessment constraints to obtain a set of verified evidence. The purification and correction module is used to determine the support type of the argument text based on the semantic-authoritative correlation between the argument text in the text framework draft and the verified evidence set, and to purify and correct the argument text based on the support type to obtain the final research report. Based on the semantic-authoritative correlation between the argumentative text in the draft text framework and the verified evidence set, determine the type of support for the argumentative text, including: Iterate through the argument text of the text framework draft and calculate the semantic similarity between the argument text and the verified evidence in the verified evidence set; The type of evidence channel used to obtain verification evidence is determined, and the authority weight value of the verification evidence is obtained based on the evidence channel type. The semantic-authority correlation degree is calculated based on the semantic similarity and authority weight values. The first K pieces of verification evidence are selected based on the similarity to form a candidate evidence set. When the number of candidate evidence in the candidate evidence set is ≥3 and the semantic-authority correlation is ≥0.8, the support type is sufficient support. When the number of candidate evidence is ≥2 and the semantic-authority correlation is in the interval [0.5, 0.8), the support type is partial support. When the candidate evidence set is empty and the semantic-authority correlation is <0.5, the support type is no support. When there are two pieces of highly similar evidence with opposite polarities in the candidate evidence set, the support type is conflicting support. The semantic-authority correlation The calculation formula is as follows: ； In the formula, For argumentative texts, This is the i-th piece of evidence in the global evidence pool; For semantic similarity; This represents the authority weight of the evidence. Similarity weights; Let c be the semantic vector of the text. For the i-th piece of evidence The semantic vector.

Images